# Scheduled denoising autoencoders

arXiv: https://arxiv.org/abs/1406.3269 · HTML: https://mlsanity.com/paper/1406.3269/

## Key facts

- Cited by 11 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 12 Jun 2014.

## Built on by (most cited first, 11 of 11)

- [Score-based Generative Modeling in Latent Space](https://mlsanity.com/paper/2106.05931/) (2021)
- [Towards the Detection of Diffusion Model Deepfakes](https://mlsanity.com/paper/2210.14571/) (2022)
- [Diffusion Model as Representation Learner](https://mlsanity.com/paper/2308.10916/) (2023)
- [SODA: Bottleneck Diffusion Models for Representation Learning](https://mlsanity.com/paper/2311.17901/) (2023)
- [Diffusion-Based Representation Learning](https://mlsanity.com/paper/2105.14257/) (2021)
- [Learning Data Representations with Joint Diffusion Models](https://mlsanity.com/paper/2301.13622/) (2023)
- [Understanding Representation Dynamics of Diffusion Models via Low-Dimensional Modeling](https://mlsanity.com/paper/2502.05743/) (2025)
- [Pre-Training Protein Encoder via Siamese Sequence-Structure Diffusion Trajectory Prediction](https://mlsanity.com/paper/2301.12068/) (2023)
- [A statistical physics framework for optimal learning](https://mlsanity.com/paper/2507.07907/) (2025)
- [Pretrained Reversible Generation as Unsupervised Visual Representation Learning](https://mlsanity.com/paper/2412.01787/) (2024)
- [Diffusion Model from Scratch](https://mlsanity.com/paper/2412.10824/) (2024)

## Abstract (arXiv)

We present a representation learning method that learns features at multiple different levels of scale. Working within the unsupervised framework of denoising autoencoders, we observe that when the input is heavily corrupted during training, the network tends to learn coarse-grained features, whereas when the input is only slightly corrupted, the network tends to learn fine-grained features. This motivates the scheduled denoising autoencoder, which starts with a high level of noise that lowers as training progresses. We find that the resulting representation yields a significant boost on a later supervised task compared to the original input, or to a standard denoising autoencoder trained at a single noise level. After supervised fine-tuning our best model achieves the lowest ever reported error on the CIFAR-10 data set among permutation-invariant methods.